Textbook2Video - 2nd Place Antler X Nvidia Hackathon

TLDR: Live Demo

Agentic LangChain pipeline that turns PDFs into animated, narrated lessons using Deepseek OCR and ElevenLabs TTS; fine-tuned Llama 3 via LoRA + 4b quantization on Brev and containerized the model for live deployment on HuggingFace.

NeuroNest - ML Defect Detection System

TLDR: Live Deployment

Production-grade defect detection system using Vision Transformers and MASK R-CNN fine-tuned on distributed A100 GPUs, achieving 98% precision. Built FastAPI REST service with PostgreSQL/Docker deployed on Huggingface; accelerated inference via TensorRT reducing latency and costs by 40%. Implemented automated CV labeling pipeline with Detectron2, CVAT, and vision LLMs, cutting manual labeling hours by 80%.

FusionTransformer - Multimodal Fall Detection

Dual-stream transformer architecture for real-time fall detection on smartwatches, fusing accelerometer and gyroscope data through Kalman filtering. Implements Squeeze-and-Excitation attention and cross-modal knowledge distillation, validated on 51-subject SmartFallMM dataset with LOSO cross-validation.

FPGA-Optimized Facial Recognition

TLDR: YouTube

Facial recognition on AMD Kria KV260 SoC achieving 99.47% accuracy with ensemble detection/landmark models; engineered zero-copy DMA + hardware-accelerated GStreamer pipeline with INT8 Vitis AI/Vivado optimizations, delivering 100x CPU speedup and 30–500+ FPS throughput.

Sensor Fusion for Human Activity Recognition

Cross-modal distillation pipeline for fall detection using accelerometer/gyroscope data with Complementary, Madgwick, Mahony, and EKF filters tuned for edge deployment and real-time responsiveness.

3D Skeleton Reconstruction from Video

Implemented GAST-NET for reconstructing 3D human skeletal joints from 2D video, pairing PyTorch-based pose estimation with computer vision preprocessing for motion analysis.

Time Series Orientation Estimation

Implemented Kalman, Extended Kalman, and complementary filters for IMU orientation estimation and sensor alignment, optimized for low-latency fall detection on embedded devices.

Smartphone-Based Fall Detection

Android fall detection app leveraging on-device IMU streams with TensorFlow Lite/ONNX models, dynamic preprocessing configs, and async sensor pipelines for reliable real-time alerts on constrained hardware.

Autonomous Chess Bot

TLDR: YouTube

Architected multi-process chess system with C++ TCP server, IPC messaging, and browser automation pipeline; containerized deployment and Stockfish orchestration deliver fully autonomous online play with 100% winrate in live games.

Real-Time Audio Transcriber

Engineered full-stack audio transcription web application with Flask WebSocket backend, React.js/Next.js frontend, hardware-accelerated Whisper AI speech-to-text, and MongoDB session storage.